Scientific Graphics Program Synthesis via Dual Self-Consistency Reinforcement Learning

2026-04-07

AI Summary

This work addresses the challenge in scientific diagram program synthesis where poor data quality and the absence of a comprehensive evaluation framework hinder models from generating executable TikZ code that is both visually accurate and structurally coherent. To overcome this, the authors propose the SciTikZ framework, which introduces SciTikZ-230K—a high-quality dataset comprising 230,000 samples—and SciTikZ-Bench, a multidimensional benchmark spanning 11 scientific disciplines. They further develop a novel dual self-consistency reinforcement learning approach, integrating an execution-driven data engine with a round-trip validation mechanism to enhance generation fidelity. The resulting SciTikZer-8B model significantly outperforms state-of-the-art models such as Gemini-2.5-Pro and Qwen3-VL in both visual fidelity and structural logic, achieving new state-of-the-art performance.

Abstract

Graphics Program Synthesis is pivotal for interpreting and editing visual data, effectively facilitating the reverse-engineering of static visuals into editable TikZ code. While TikZ is the de facto standard for scientific schematics due to its programmatic flexibility, its requirement for rigorous spatial precision presents a significant challenge for Multimodal Large Language Models. Progress is currently stifled by two primary gaps: (1) Data Quality Gap: existing image-TikZ corpora often lack strict executability and reliable visual alignment; (2) Evaluation Gap: a lack of benchmarks for both structural and visual fidelity. To address these, we present a closed-loop framework featuring: SciTikZ-230K, a large-scale, high-quality dataset from our Execution-Centric Data Engine covering 11 diverse scientific disciplines; SciTikZ-Bench, a multifaceted benchmark spanning from basic geometric constructs to intricate hierarchical schematics to evaluate both visual fidelity and structural logic. To further broaden the scope of visual-code optimization methodology, we introduce a novel Dual Self-Consistency Reinforcement Learning optimization paradigm, which utilizes Round-Trip Verification to penalize degenerate code and boost overall self-consistency. Empowered by these, our trained model SciTikZer-8B achieves state-of-the-art performance, consistently outperforming proprietary giants like Gemini-2.5-Pro and massive models like Qwen3-VL-235B-A22B-Instruct.

Problem

• Graphics Program Synthesis
• Data Quality Gap
• Evaluation Gap
• TikZ
• Scientific Schematics

Innovation

• Graphics Program Synthesis
• Dual Self-Consistency Reinforcement Learning
• Execution-Centric Data Engine
• SciTikZ-Bench
• Round-Trip Verification